Ollama — Product development¶

Development line: project:ollama · thread product-development
Last event: 2026-01-22 · 2 dated since 2025-05-19 · Researched: 2026-09-04 · confidence: medium

What it is¶

Ollama is a local runtime for developers running language, vision, and embedding models on their own hardware or through a compatible API.

• Model runner: downloads models and provides an interactive CLI, an HTTP API, and Python and JavaScript libraries.
• Vision processor: handles images alongside text prompts.

Speed, context size, and concurrency depend on available RAM and VRAM.

A practical alternative to cloud inference for local development and controlled data placement.

Development line¶

• 2025-05-19 — Ollama published an update about multimodal models. On 2025-05-19, Ollama linked to its official blog page for multimodal models and to its download page without details on specific models or hardware requirements.
• 2026-01-22 — Ollama shared a software-release update. On 2026-01-22, Ollama directed users to the GitHub releases page without naming the version or change list.

What changed¶

• 2025-05-19 — Ollama moved vision models to a new multimodal engine and added support for Llama 4, Gemma 3, Qwen 2.5 VL, and Mistral Small 3.1.
• 2026-01-22 — A link to GitHub Releases confirms the release channel, but the URL omits a version tag, so the release contents cannot be reconstructed reliably.

How to use this¶

As of 2025-05-19, assess the official multimodal workflow when choosing local models. As of 2026-01-22, check the project's GitHub release notes before upgrading.

1. Install Ollama on macOS, Windows, or Linux. — https://docs.ollama.com/quickstart
2. Run a model in the terminal, such as ollama run gemma4, to test the first session. — https://docs.ollama.com/quickstart
3. Send requests from applications to the local API at http://localhost:11434/api, or use the official Python and JavaScript libraries. — https://docs.ollama.com/api/introduction
4. Pass an image alongside text in the CLI or in the images field of the API request for vision tasks. — https://docs.ollama.com/capabilities/vision

Best practices¶

• Check model placement with ollama ps to verify whether it runs on GPU, CPU, or in mixed mode. — https://docs.ollama.com/faq
• Match OLLAMA_CONTEXT_LENGTH and concurrency to available memory: consumption grows with parallel requests multiplied by context size. — https://docs.ollama.com/faq
• Keep the service on the default loopback interface; change OLLAMA_HOST deliberately and use a reverse proxy when exposing it. — https://docs.ollama.com/faq
• Disable cloud features with OLLAMA_NO_CLOUD=1 or disable_ollama_cloud for an entirely local setup. — https://docs.ollama.com/faq

Superseded by this¶

• 2025-05-15 — The older integration built on shared multimodal code on llama.cpp was replaced by isolated model implementations in the new engine.
• 2026-01-20 — Image generation restricted to macOS was marked experimental; Windows and Linux were not supported yet.

Still unknown¶

• The GitHub Releases link in the 2026-01-22 event points to a mutable release feed without a tag, so we cannot tie a specific version or changelog to that date.
• Current documentation includes cloud models, but the 2 dated links do not show whether cloud execution was part of the original announcement.

Sources¶

Agent brief¶

• Subject: project:ollama, thread product-development, 2 dated events 2025-05-19 → 2026-01-22.
• Practical note: As of 2025-05-19, assess Ollama's official multimodal workflow when selecting local models; as of 2026-01-22, review the project's GitHub release notes before upgrading.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.